26.04.2023 – Gastvortrag: Dr. Katerina Kalouli (LMU München)
Marrying up symbolic reasoning and LLMs: the promise of hybrid systems
Abstract: The advances and success of Large Language Models (LLMs) have revolutionised many fields of NLP. Massive deep models are arguably performing similar to humans. However, a closer look at the performance and inner workings of these models reveals their generalisation difficulties on unseen data, their inability to perform human-level reasoning and their black-box nature that makes it hard to improve them in the right way. On the other hand, traditional rule-based systems can handle many of these challenges in a precise and explainable way. These findings have opened the way for more hybrid approaches, combining the strengths of symbolic reasoning and LLMs. In this talk, I will present one such approach implemented for the task of Natural Language Inference (NLI) and show how the two worlds can be married up to a harmonious combination. I will also present some findings on the inner workings of LLMs, which additionally confirm the necessity for a hybrid approach.